Combined eye activity measures accurately estimate changes in sustained visual task performance

Five concurrent eye activity measures were used to model fatigue-related ch anges in performance during a visual compensatory tracking task. Nine parti cipants demonstrated considerable variations in performance level during tw o 53-min testing sessions in which continuous video-based eye activity meas ures were obtained. Using a trackball, participants were required to maneuv er a target disk (destabilized by pseudorandom wind forces) within the cent er of an annulus on a CRT display. Mean tracking performance as a function of time across 18 sessions demonstrated a monotonic increase in error from 0 to 11 min, and a performance plateau thereafter. Individual performance f luctuated widely around this trend - with an average root mean square (RMS) error of 2.3 disk radii. For each participant, moving estimates of blink d uration and frequency, fixation dwell time and frequency, and mean pupil di ameter were analyzed using non-linear regression and artificial neural netw ork techniques. Individual models were derived using eye and performance da ta from one session and cross-validated on data from a second session run o n a different day. A general regression model (based only on fixation dwell time and frequency) trained on data from both sessions from all participan ts produced a correlation of estimated to actual tracking performance of R = 0.68 and an RMS error of 1.55 (S.D. = 0.26) disk radii. Individual non-li near regression models containing a general linear model term produced the cross-session correlations of estimated to actual tracking performance or R = 0.67. individualized neural network models derived from the data of both experimental sessions produced the lowest RMS error (mean = 1.23 disk radi i. S.D. = 0.13) and highest correlation (R = 0.82) between eye activity-bas ed estimates and actual tracking performance. Results suggest that informat ion from multiple eye measures may be combined to produce accurate individu alized real-time estimates of sub-minute scale performance changes during s ustained tasks. (C) 2000 Published by Elsevier Science B.V.